Re: Optimizing Across && And ||

Preston Briggs <preston@tera.com> noted that his tests of compiler
optimizations showed disappointing results, and gave some explanations
he'd heard for this. Let me add to the possible explanations one more
very important one: A friend pointed out to me that these days compilers
mostly live or die by the SPEC benchmark suite. Optimizations which don't
improve most of the benchmarks in this suite significantly are considered
failures under this metric, due to the averaging procedure. This is very
bad, inasmuch as many of the SPEC tests are bottlenecked by simple, tight
numerical loops -- optimizations such as global value propagation are
useless in these cases, whereas optimizations such as instruction
scheduling are extremely important.


It's an interesting metric to look just at the GCC SPEC benchmark -- this
exercises quite a bit more compiler optimization than most of the rest of
SPEC.


Bart Massey
bart@cs.uoregon.edu
--

Return to the comp.compilers page.
Search the comp.compilers archives again.